2.2.4. Imagery Feature Extraction

The construction of the categorical dataset requires the spectral signature, index feature calculation, gray-level co-occurrence matrix extraction, and feature fusion [41]. The spectral signature is the foremost feature in remote-sensing image classification [42], while the index feature can effectively enhance image classification accuracy. Index features include Normalized Difference Vegetation Index (NDVI) [43], Normalized Difference Water Index (NDWI) [44], and Normalized Difference Built-up Index (NDBI) [45,46]. Besides, this study adopts the Combined Mangrove Recognition Index (CMRI), which can enhance the discrimination between mangrove forests and nonmangrove vegetations [47] (Table 4).

The existing studies suggest that the gray-level co-occurrence matrix (GLCM) can effectively enhance the classification accuracy of various land-use types and diminish the classification errors due to similar spectral signatures [48]. Combined with the test, the study sets the window size of statistical pixels as 9 × 9 while extracting textures and selects the value of grayscale quantization level as 32 to calculate six textural features of images using the GLCM (Table 4).

After the calculations above, this step adopted the integrating the multifeatures method [49], including spectral features, texture features (GLCM), and index features. Each feature obtained in this study corresponds to a layer, and the method of layer-overlay

is used for feature fusion (Figure 3). The integrating multifeatures method can significantly enrich the information content of remote sensing data, and one of the most commonly used methods to quantify the importance of features is decision trees.

**Table 4.** The characteristic attributes involved in classification.


**Figure 3.** Multifeature fusion diagram (take Hailing Island as an example).

2.2.5. Random Forest Classification

Random forest (RF) is an integrated learning technology that can generate a large number of decision trees for the training, calculation, and classification of samples. The bagging method is adopted in RF to generate independent, identically distributed training sample sets for each decision tree, and the final classification result of the RF depends on the voting of all decision trees.

The RF algorithm is amplified as follows: (1) obtain *N* (*N* is a random positive number) training sample sets from a large number of original samples, by drawing with replacement *N* times; (2) select *m* (*m* is a random positive number) classification features randomly from the total features in each sample set; (3) divide the nodes of the decision trees by complete segmentation methodology, and then build a great number of decision trees. After completing the classification for each decision tree, the classes of new samples are determined by a majority vote according to the classification results of the decision trees [50].

The main idea of using the random forest to measure the importance of features is to evaluate the contribution of each feature in each decision tree, and to calculate the average values. Subsequently, the contribution value of features can be compared.

#### 2.2.6. Visual Modification

This study involves various classification types, with the existence of the phenomenon of "different objects with the same spectrum". This leads to unavoidable ineffectiveness when distinguishing surface features. Therefore, in combination with the auxiliary data, the classification maps generated by RF were visually modified to enhance the classification accuracies.
